v10.0.0

v10.0.0 — Rebuilt Interface & CPU-Only Mode

A major release: every tab is rebuilt as a modern page, the program now installs and runs without an NVIDIA GPU, the AI stack moves to PyTorch 2.14 and transformers 5, and dozens of fixes make database builds, OCR, search and speech more reliable.

🖥️ A rebuilt interface

• Every tab, the Ask Jeeves window and the Chat Backend Settings dialog were rebuilt as modern pages that share one theme and one set of tooltips and menus (0dd6d72)
• The Models tab shows each embedding model as a card with its specs, benchmark scores, the precision it will run in on your computer and its own Download button, plus search, filters, sorting and a compact list view (ec0231a)
• The Settings tab fits on one screen, saves each change as you make it and checks values as you type (1a4b894)
• The Tools tab shows each tool's progress and results inside the tab, including a page-by-page OCR progress bar, and a sortable GPU list replaces the old chart (6251eb0)
• The Create Database tab lets you add and review files inside the tab, checks the database name as you type, runs the PDF check in the background and shows each build stage live with a log (e4bb37c)
• Tokenization now appears as its own build step between Splitting and Embedding (cb7b528)
• The Manage Databases tab lists every database with its size, file and chunk counts and backup status, opens or reveals the original files, and deletes, backs up or restores a database after a confirmation in the tab; its button for searching a database is now called Query (a791671)
• The Query Database tab streams answers into cards with their sources, scores and page numbers, shows each passage in Chunks only mode, and handles missing keys, speech, dictation and model ejecting in the tab instead of pop-ups (7ce207a)
• Ask Jeeves is now a chat-style window that keeps its familiar picture and shows sources, read-aloud and poems inline (d0d847d)
• Chat Backend Settings lists each backend with its status, saves changes as you make them, keeps keys masked, adds a key field for MiniMax and can test the connection to ChatGPT, LM Studio and Kobold (5e421a7)
• The user guide and the Ask Jeeves knowledge base were updated for every new tab, model and setting (3be4010)

💻 CPU-only mode

• The program now runs on computers without a supported NVIDIA GPU: the installer sets up the CPU version automatically (or when run with --force-cpu), and at startup a GPU too old for the bundled PyTorch is skipped so the program runs on the CPU instead (c2313fc)
• File transcription and InternVL vision models work on the CPU, Half precision no longer blocks CPU database builds, and a build with images is refused up front if the chosen vision model needs a GPU, instead of silently skipping the images (844208e)
• CPU installs offer only the chat, vision and embedding models that performed well in CPU benchmarks, and a saved GPU-only choice switches to Liquid-VL 480M or is refused with a clear message (d858cfd)
• CPU installs show speed notes and LM Studio tips, a failed text-to-speech run is reported in a dialog, and the README and guide explain CPU-only mode (f1373c3)
• The Models tab's CPU build-time estimates now cover the new embedding models (a54b2ac)
• On a GPU computer, choosing CPU for database creation now asks for confirmation instead of always being refused (47694a1)

✨ New models

• New embedding model: Jasper-Token-Compression-600M (b81783c)
• New embedding model: Yuan-embedding-2.0-en, a 596M English retrieval model (6ca0284)
• New embedding model: F2LLM-v2-1.7B, especially strong on code (b0e51f5)
• New embedding model: F2LLM-v2-4B, for GPU installs (3ab12ff)
• New embedding model: geevec-embeddings-1.0-lite (5f6514a)
• Each embedding model now lists its multilingual, English, legal and code retrieval benchmark scores instead of a rough rank (fdac033)
• Two new Whisper models in the Transcribe File tool, Whisper large-v3 turbo and Distil Whisper large-v3.5, each in float32, bfloat16 and float16 (f5f99d6)
• Kokoro text-to-speech is now available throughout the program with voice and speed settings, and every install downloads it (e8317cf)

🔎 Better search results

• EmbeddingGemma now uses Google's query and document prompts, which markedly improves search results; existing databases get most of the gain without being rebuilt (bc505e9)
• EmbeddingGemma now uses its full 2,048-token window instead of ignoring text past the first 512 tokens of each chunk (3d0acf0)
• The e5-v2 models and Free Law's ModernBERT 512 now use the search prefixes they were trained with, which improves retrieval; recreate databases built with these models to benefit (6beaacb)
• Instruction-tuned embedding models now receive their search instruction in the exact format their authors specify (f00c215)
• Each embedding model's type is now read from its own folder name, so an install path containing words like bge or qwen no longer gives models the wrong prompt or length, which could make database creation fail (1ae5756)
• Searches limited to images, audio or documents keep looking until they find enough matching chunks, instead of often coming back empty (e1db3eb)
• The default similarity is now 0.5 instead of 0.8, which often filtered out every chunk, and the Similarity tooltip and guide now explain that higher values return fewer, closer matches (a8049a0)
• EmbeddingGemma and Harrier 270M no longer crash on documents containing text that looks like their internal special tokens (0c49b99)
• EmbeddingGemma and Harrier 270M stay in float32 on GPUs older than the RTX 30 series, where Half precision made them produce unusable vectors, and the Half precision tooltip now says so (977605b)

⚡ Faster and lighter

• The main window now opens in about 3 seconds instead of 8 in testing and uses about 450 MB less memory, because each feature loads its libraries the first time you use it (9a67ce0)
• Background helper processes no longer load the whole program, so they start faster and read-aloud begins speaking sooner (3bdd6cf)
• Creating databases is up to 13% faster with EmbeddingGemma and slightly faster with the Qwen3-based models, with identical results (34010fd)
• Retuned build batch sizes make EmbeddingGemma faster on typical chunks without running out of memory on long ones, and Harrier 270M faster on longer chunks with less VRAM (affa5b8)

🛡️ Safer builds, backups and downloads

• Files a build leaves out, such as failed extractions, unreadable images or files with no text, are now listed in the console and the completion message, and no longer show as embedded in the Manage tab (0363f11)
• A crashed build now removes its partial database, temporary files are deleted however a build ends, and a crash after the database is finished no longer reports a failure (3d6b956)
• Pressing Cancel in the last moments of a build keeps the finished database instead of deleting it, and an early Cancel stops the build before it starts (3c024ed)
• If a worker process crashes during text extraction, the extraction step is retried instead of finishing with no documents (fff3fa5)
• Create Database waits until files finish being added to the list, so files added late are no longer silently left out (99a8d10)
• Transcribing and building a database no longer overlap, which could delete a new transcript without adding it (12d1da4)
• Closing the program no longer deletes transcripts that haven't been built into a database yet, and recordings with the same name no longer overwrite each other's transcripts (03977ee)
• Restore Databases copies the backup first and swaps it in only after the copy succeeds, refuses when there is no backup, and shows the real error instead of reporting success (d6dcbff)
• After a restore, settings entries for databases the backup doesn't contain are removed, and the database to search is cleared if it was one of them (80ec42b)
• Closing the window during a backup or restore now asks you to wait until it finishes, instead of cutting the copy off partway (1404d3f)
• Database names are capped to stay within Windows' 260-character path limit, and the installer warns when the install folder's path is long (73b1b5b)
• Database builds no longer crash on non-ASCII file names, and setup and the pre-build check warn when the install folder's path has non-ASCII characters, which the database library can't handle (4bbbd72)
• A model download cut off by closing the program or a crash is no longer shown as downloaded or offered in the Create Database model list (423e96d)
• Downloads now always report when they finish; before, the Kokoro download's result was often lost and the program could crash (7ade040)

📄 Reading documents and images

• Emails (.eml) no longer have their text indexed twice or keep HTML tags, and emails with an unknown character set are no longer dropped (fe955f5)
• UTF-16 text, Markdown, CSV and HTML files are read correctly instead of as gibberish, and HTML files honor their declared character set (4ae1afb)
• Footnotes and endnotes in Word (.docx) and RTF files are now included (9ef799c)
• RTF files that contain emoji are no longer skipped (e0e59ce)
• Vision models now see transparent and 16-bit images properly instead of describing them as blank (f7e278b)
• Qwen-VL vision models keep their entire image encoder in full bfloat16 precision, as Unsloth's dynamic quants do, using about 0.35–0.6 GB more VRAM (31780f7)

🔍 OCR

• OCR reads multi-column pages column by column instead of mixing lines across columns, while single-column pages and upside-down scans keep their normal order (5a255fd)
• Sideways scans are detected correctly, and their searchable text layer lines up exactly with the page (2110ee8)
• The Tesseract engine no longer doubles the text on pages that already have a text layer; like RapidOCR, it copies those pages through unchanged (bf618cc)
• OCR'd PDFs keep their bookmarks, page labels and links (0f7daf6)
• The database OCR check no longer keeps flagging already-OCR'd PDFs that contain photos or pages with only a few words (c057938)
• PDFs whose text extracts as gibberish because of unusual fonts are now flagged for OCR, and once OCR'd, the readable OCR text is used for those pages (ba7c14f)
• OCR model downloads are checked against known checksums and re-downloaded if corrupt, and inconsistent OCR engine output is reported instead of being silently cut short (1bfae2d)

🔊 Speech

• Kokoro reads multi-line answers in full, and the text being read can no longer be run as a command on your computer (798b1be)
• ChatTTS works again on GPU or CPU, no longer garbles speech, and speaks numbers and symbols instead of dropping them (273bb52)
• Read-aloud no longer splits numbers like 2.5 or citations like O.C.G.A. in the middle (1f2a935)
• Kyutai Pocket text-to-speech no longer fails to start (3c1fbfd)

💬 Chat

• Switching local chat models mid-session no longer re-enables Submit too early, blanks the chat history or blends answers together (2cfe8a4)
• Closing the program during a Chunks only search no longer freezes it for up to two minutes (1920c84)
• The token counts under local-model answers now include the full prompt the model actually receives, and the too-long check uses the same count (fc2a857)

⚙️ Settings

• One out-of-range value, such as Contexts or Chunk Size, no longer resets the whole database section of your settings; every valid value is kept (265ad9e)
• config.yaml and model settings files are read as UTF-8 everywhere, so non-English characters no longer break local models, Ask Jeeves or database creation (bf25db6)
• Fresh installs start from current defaults: similarity 0.5, the local Kokoro voice, the Liquid-VL 480M vision model and gpt-5.4-mini for ChatGPT, with no leftover entries for removed features (4f3b509)

🕸️ Documentation scraper

• Pages whose addresses differ only by upper- or lowercase letters, such as torch.Tensor and torch.tensor, are now both saved (477ce77)

📦 Installation and dependencies

• GPU installs move to PyTorch 2.14.1 with CUDA 13.0, which needs NVIDIA driver 580 or newer and a GTX 16-series, RTX 20-series or newer card; older cards get the CPU version, and temporary workarounds for problems this PyTorch release fixes were removed (033999d)
• Python 3.14 is now supported (3.11 through 3.14), and Tesseract is updated to 5.5.2 (fac359f)
• Moved to transformers 5 and sentence-transformers 6, with databases, chat answers and transcripts coming out the same as before (d25aed8)
• The transcription libraries are pinned to matching versions, so fresh installs no longer get a ctranslate2 too old for whisper-s2t-reborn (759e7eb)
• Python 3.11 installs no longer crash at startup because of a mismatched torchvision (d73b16a)
• The installer no longer crashes on slower connections while downloading the large PyTorch package (b7b0f83)
• Pandoc is no longer required, and the unstructured library plus the 23 packages only it needed are gone; the program reads every file type with its own loaders (29e5781)
• flash-attn, xformers and matplotlib are no longer installed, since nothing uses them anymore (929ce70)
• Installing the ChatTTS fix no longer modifies uv's shared package cache (615a3a0)

🔧 Housekeeping

• Removed unused leftovers: the transcribe sections in config.yaml, unused Whisper model fields, a dead model rule, the old Tokenizer folder and an unused import, and fixed a Python warning from the Google TTS code (197ce3f)